BanglaFinDoc

Description

A semi handwritten financial document image dataset containing 334 Bangla financial document images with field level annotations. It includes 2001 rectangular bounding boxes of 9 kinds of field types. The fields are used to specify the vendor name, vendor date, vendor serial number, document type, goods description, quantity, unit price, total value, and truck number. The annotations are available as a Label Studio JSON export. Every record points to a single image and provides the top, left, bottom and right coordinates of each region annotated and also the region category. Such annotations can be used for training and testing models for field detection as well as for cropping specific fields prior to optical character recognition (OCR). Only the field labels and the bounding boxes are included in the JSON. There is a need for a separate set of reference transcriptions for evaluating the accuracy of text recognition. The collection has varying levels of field coverage. The most common label is Goods description, with 335 areas annotated. The document type is known in 20 regions and the truck number only in one of these. When assessing the extraction performance, the number of examples per field should be reported.

Institutions

• Daffodil International University

  Dhaka Division, Dhaka

Categories

Licence